There is no point multiplication operation on secure elliptic curves*, there is only scalar multiplication apart from the point addition of the group.
Scalar multiplication
Scalar multiplication with a scalar $t$ is adding a point $P$ it self $t$-times
$$ [t]P : = \underbrace{P + P + \cdots + P}_{t-times}$$ and it is well defined operation since the group is operation is the addditon of the points. Scalar multiplication is a natural writing of the adding $t$-times.
If you want to execute a Diffie-Hellman here it is;
Elliptic Curve Diffie-Hellman Key Exchange (ECDH)
ECDH is executed with the parties' public keys. Let say Alice has secret key $a$ and public key $A = [a]G$ and Bob has secret key $b$ with public key $B = [b]G$ with the base point $G$ defined by the standards.
Then both parties can agree on the key as;
- Alice: get Bob's public key $B$ and multiply it with secret key $a$; $S = [a]B=[ab]G $
- Bob: get Alice's public key $A$ and multiply it with secret key $b$; $[b]A=[ab]G = S$
So both arrived at the same point $S$, now they can use the $x$-corrdinate point $S$ with a Key Derivation Function (KDF) like HKDf to derive a symmetric key.
Ephemeral
Originally, Diffie-Hellman is defined as ephemeral, that is for each key generation, Alice and Bob generates new uniform random points by selecting a uniformly random $a$ and $b$ per run of the protocol. This provides forward secrecy if you delete the keys after decryption.
If you don't generate new uniform random keys, it is called the static DHKE. Therefore you will see around static-static, ephemeral-static, static-ephemeral, and ephemeral-ephemeral usages of the DHKE on around. Hopefully, the static case is no longer available in TLS 1.3.
MITM attack
DHKE is vulnerable to a man-in-the-middle attack especially the ephemeral case since the public points are changing all the time. A signature must be used to prevent this attack.
* As noted by CodeinChaos in the comments we can define a multiplication on curves with $[a]G \cdot [b]G = [ab]G$. This, clearly, have commutativity, associativity, identity, and distribution law. The inverse depends on the order of the $G$. This multiplication, however, is exactly the Computational Diffie-Hellman (CDH) problem. If you solve the Discrete Log (DL) problem on the curve then you can solve the CHD, the reverse is not known in the general case.
Since we required that the DL (or CDH) must be infeasible for security, this multiplication operation is not a feasible operation on the secure curves.